Networks, Crowds & Markets Final Paper

Networks, Crowds and Markets
Final Paper
May 9, 2012
Stacy Feldman
Jason Rawlins
Katrin Warren
Marina Zavelevich (MarZ)

ABSTRACT
Traditional forms of media advertising use various forms of ROI to justify media spend
and track its effectiveness in changing consumer behavior and driving purchases. In the last few
years, marketing efforts have focused on the emergence of social media as an important tool to
communicate brand messages, resulting in the rise of “network-based marketing,” which seeks
to use the links between consumers to increase sales. Moreover, a number of different forms of
measurement have evolved to track the links between consumers. These services give scores
to individuals, or to specific business accounts, to represent their respective level of influence
within a given community.
In this paper we will examine the various measures of influence in social media. We
will evaluate the strengths and limitations of the existing models by comparing the rankings of
key individuals or accounts, as the case may be that score highest in specific models across a
collection of social media influence measuring sites. In reviewing academic studies, we will
analyze whether the set of key individuals‟ rankings are in accordance with expectations set by
models of these sites; for instance, whether a celebrity or news outlet scores as high as
expected in relation to one another on a certain influence measure service. Then, detect why
their rankings may vary across the different services; which model attributes contribute to that
ranking being higher, lower, or the same when evaluated through a different site.
The comparison of account scores to one another and across services will be scaled
against the inherent level of influence in their relative fields; consequently, comparing the
relativity of magnitude in determining the reach or penetration of a particular influencer‟s
message. For instance, a high level industry executive‟s influence would be measured against
the magnitude of the industry‟s professionals‟ presence in social media outlets, versus a pop
singers influence over an appropriate youth demographic.
Through these observations we will study the validity of each model‟s definition of
influence via its measure metrics. Beyond an individual‟s reach or popularity how do these

models track ability to cause change in behavior of others? And for qualified observations, how
can influence be inferred through scope and spread of messages or ideas (a.k.a. change of
behavior) funneled through social media outlets? Additionally, by tracing whether the effected
change is predictable for the network subset due to shared predisposition or whether it is out of
the otherwise standard behavioral norm, we will attempt to make the distinction between change
of behavior as a cause of external forces versus a result of homophily – drawing a bridge
between homophily as a symptom of correlation and potential inferences of causation as a sign
of influence.
INTRODUCTION
Over the course of the last decade, social media has transformed the way individuals
and businesses interact with one another. The traditional marketing and advertising methods
have been modified to accommodate consumers in the digital landscape we now inhabit. The
rise of the Internet has facilitated in accelerating the time it takes for a message to diffuse
throughout the target market and beyond. It has also allowed both consumers and producers to
engage in a free-form dialog that previously existed in focus groups, usually managed by a third
party agency. In this new world, companies can speak directly to their consumers and the
consumers can reply, in real time, to profess their adoration or discontent for a product or
service.
Consumers can do more than speak with the business; they can also engage others,
who may or may not endorse the business themselves, in discourse around the pros and cons
that the business brings to the table. These discussions all take place in a public space, where
others in the online community can add their experiences and opinions on the business.
Through these discussions, consumers build and earn trust in one another based on the
value of the content that is posted. Certain consumers in the online community gain a
reputation for the quality of the content they share with the masses, and the measurement of

their reach and influence is typically qualified in the number of people who subscribe to their
blog posts or the number of followers on Twitter or friends on Facebook. This qualitative look at
these hubs of information does prove important, but the big mystery lies in how we can quantify
reach and influence through social media impressions.
Businesses must be players in social media in order to survive in today's market. They
are working tirelessly to make sure they have significant online presence on the major social
media networking sites like Facebook, Twitter and LinkedIn. Due to the popularity of certain
social media platforms amongst millions of users, brands have recognized an opportunity to
leverage these sites and their users to target their brand advocates and further their business
goals. Brands are targeting social media influencers (SMIs) across various platforms to impact
behavior, generating viral word of mouth to promote their products, messages, or services.
Traditional marketing and advertising methods have metrics and scoring systems that are
accepted as industry standards to determine ROI. Standardized methods and scoring systems
for measuring ROI within social media networks, comparable to these traditional metrics, are yet
to be universally accepted. Moreover, companies are carving out market niches by establishing
proprietary metrics.
We will focus on several social media networks that measure a user‟s digital footprint
across social media platforms.
THE IMPORTANCE AND DEFINITION OF INFLUENCE IN SOCIAL MEDIA
The industry has begun quantifying influence in social media. Social network analytics
provide underlying network structure data that details insights on strong and weak ties, hubs,
local bridges and structural holes. This data is leveraged to „score users‟ influence across a
social network by measuring the level of engagement that users have with the influencer‟s
content. One difficulty in accurately measuring influence is that the same factors that go into
measuring it can also be used to claim correlation or environmental externalities as the reason

for engagement in a social network. In order to distinguish between the three, we must first
define influence as it relates to social media.
Here, we define social influence as a node‟s ability to induce a response from or diffuse
information through the node‟s strong and weak ties. If the node is influential enough, its
message can cascade across a network and go „viral‟. At the surface level, a node is said to be
influential if it has multiple followers/friends, receives likes/retweets, or is re-blogged and pointed
back to as a point of reference on a given topic. The tools currently available are limited in their
ability to accurately score the reach of influence. Furthermore, due to crude keyword
recognition technologies, it is challenging to distinguish a topic that is initiated by a user versus
a topic with which a user is involuntarily associated.
It is also difficult to ascertain precisely what makes a topic spread through a network, but
a high degree of correlation can imply potential causality.1
Budak et al. differentiate between
correlated and uncorrelated structural trends within online social networks. Correlated trends
involve niche topics that a small cluster of nodes discuss, whereas uncorrelated trends include
topics that are of interest to a large number of people and many of whom are unrelated to one
another.2
Budak et al. conducted an experiment comprised of a network of five hundred nodes
in which fifty varying topics were introduced. Not only did the topics vary, but the nodes‟
behaviors also differed based on their preferred interest in some topics over others. Further,
some nodes are more easily influenced by their local neighbors or by external forces.3
Budak et al. viewed five hundred hashtags on Twitter and segregated them into seven
different topic categories: politics, technology, celebrities, games, idioms, movies, and music.
They found that political hashtags have a highly correlated importance that may suggest
homophily, where people share information with liked minded friends. Differently, for a category
such as idioms, users interested in this topic will follow tweets using that specific hashtag as

opposed to individual users with whom they are friends. Therefore, users are not opting to
follow “friend” or someone to whom they are socially tied; rather, the user chooses to follow a
tweet that mentions a specific hashtag. This behavior seems to be specific to Twitter because
the platform, by nature, broadcasts the tweets of every user where all cited hashtags are
searchable.4
This is further evidence that environmental influence may play an important role in
identifying the effects of homophily versus the effects of influence. Before ascertaining
influence we must first control for homophily as current metrics will show homophily as having
the same network correlations as influence. To further isolate the effects of influence, we must
determine the difference between correlation and causation.
Anagnostopoulos et al. sought to define correlation, versus causation, by establishing
three distinct models with the following conditions: (i) they isolated a network from the effects of
internal influence, then (ii) they isolated a network from the effects of external influence, and
ultimately (iii) they rendered a network independent of both, where externalities in the network
moved these nodes to the same actions. These models can also be referred to as the
homophily model, the influence model, and the confounding model, respectively. The
researchers hypothesized that the action one takes is not as a result of an influencer, and
therefore adjusting the timing of the delivery of the influencer‟s message would not reveal a
change in behavior.5
They then prove that influence is not a factor contributing to the presence
of correlation in the homophily and confounding models. Restated to fit the nature of our paper,
social correlation plays a bigger role in actions throughout a network than does causation.
The real world data used in these tests referenced above came from Flickr and the
action of tagging an image is the metric used to evaluate the effectiveness of influence.
Tagging is the action of marking a picture with a phrase that identifies or characterizes the
image. Sixteen months of tagging data, spanned across approximately 800K users, was
whittled down to approximately 340K users who actively tag photos. The tags themselves were

analyzed and from a collection of 10K, the researchers decided to focus on 1,700 tags across
various topics that exhibited various growth patterns (i.e. current events which can have volatile,
unexpected spikes and dips as opposed to holidays which come with anticipated expectations
for tagging).6
The first series of tests were conducted as simulations to prove that the models could
actually measure correlation. These tests mimicked the growth rates of new users and new
connections in the Flickr data set to its models, so that the tagging patterns in its tests
accurately reflected the rate of tagging in the real data. Once ensuring the integrity of the test
data, the models generated networks where tagging spread with no correlation, where tagging
spread with influence, and where tagging spread with correlation (minus influence).7
After
establishing these networks, another series of tests were run to distinguish the presence of
influence from the presence of correlation; these tests are referred to as the Shuffle test and the
Edge-Reversal test (inspired by obesity studies by N.A. Christakis and J.H. Fowler)8
. The
Shuffle test shows us that if influence is not present, the timing of a tagging should be
independent of the tagging of others in the network, even in the instances where homophily is
present. The Edge-Reversal test is a regression model that reverses the direction of all the
edges in the network to see if correlation is still present. One should expect this to be the case
where, if homophily is present, friends will still tag one another‟s photos without, where any
friend is equally likely to be the first or the last to tag and item. In the case where tagging occurs
as a result of influence, the edge-reversal should show a drastic change in correlation9
.
The outcomes of these tests are charted in the accompanying graphs (Exhibit A)10
.
Figure 5 shows the results of shuffling the timing of tags in the correlation model, and the results
prove that the original and shuffled tags follow the same path, therefore indicating a lack of
influence. Figure 4 shows the results of shuffling the timing of tags in the influence model, and
as the name indicates, influence is clearly present; the results prove that when the timing of the

tags is shuffled, tagging does not occur at the same rate or in the same path as the original data
set. Figures 7 and 6 (Exhibit B)11
give us the same results as the influence tests did in the
graphs depicted in Figures 5 and 4, respectively. The graphs validate the assumption that
social correlation is less present in both the shuffle and edge-reversal tests of the influence
model, than in the correlation model. The results of the test data, however, were not consistent
with the data retrieved from the Flickr dataset. The Flickr data displayed shuffle times and
reversed edges that followed the original data paths very closely (Exhibit C)12
. We can infer that
the influence we saw in Figures 4 and 6 of the test data is not present in a real world scenario.
This can mean a number of things, specifically, users tag images but they rarely re-use the tags
previously set by influencers in the network.
In examining Flickr as a social network, Anagnostopoulos proves that influence has very
little effect in the spread of photo tagging. Although the experiences on Facebook and Twitter
are different than Flickr, we can use this study to say the same for other social networks and the
idea of “liking” or “retweeting.” Correlation has a deeper effect on nodes of a network than does
causation. Research is still ongoing on this topic because identifying causation as a catalyst to
action still cannot be quantified to the point of creating a standard measurement of influence.
LEADING SOCIAL MEDIA METRICS
Currently, there are a variety of measures of Social Media Influence. Many of these
influence measures use similar inputs; yet place different levels of importance on them as they
are placed into algorithmic or weighted models. Ultimately, each measure positions its results
differently as a means to market its service to social and business communities at large. This
section will identify and outline a variety of different social media influence metrics and will
broadly discuss what each measurement claims to do, how scores are determined, the degree
to which you and/or your peers effect your score, how often your score is calculated, and

limitations of the model. (An accompanying comprehensive chart of these findings is included
as Exhibit D.) In light of the various social medial influence measures, businesses must
determine which scores suggest the ability of a person to influence others, and take into
consideration which scores are most important given their marketing objective, realizing they
may not correspond with the “highest score.”
Klout:13
Klout measures influence based on a user‟s ability to drive action in his/her
social networks. Klout collects publically available data as means for input to a user‟s score.
Users can control the data available to Klout by changing their privacy settings on individual
networks. These inputs include Twitter (retweets and mentions), Facebook (comments, wall-
posts, likes), LinkedIn (comments, likes), Foursquare (tips, to-do‟s, done‟s), and Google+
(comments, reshares, +1s). The company also uses blogs and activity on Facebook pages,
YouTube, Instagram, Tumblr, Blogger, Wordpress, Last.fm, and Flicker. In addition, the
company weights these measures not just on level of activity, but on how influential a person is,
meaning how much engagement do they generate. They focus on three key measures: True
Reach (how many people a user influences), Amplification (how much a user influences
him/her), and Network Score (how influential his/her peers are).
Klout scores are on a scale of 1-100 and as a user‟s score increases, it is exponentially
harder to increase a user‟s score, where the average is in the 20s vs. the 50s. Klout
recalculates scores every day. The company suggests users can boost their scores in three
main ways, by (i) creating content worth sharing, (ii) starting discussions, and (iii) registering
and connecting other social network feeds. Furthermore, Klout helps companies reward users
for their influence on a given topic, providing them not only with a higher influence rating on a
given topic, but also connecting users with companies and brands who may offer them
exclusive sneak peeks, service upgrades and free goods to try out.14
However, as one critic
points out, “I am not aware of any credible studies that demonstrate ANY tie between Klout

score and the observation of some desired behavior BEYOND a mere retweet. There is no
evidence that a message from an influencer has any impact on trial, usage, churn or customer
satisfaction.”15
Thus, while testing could be done, nothing conclusive has come to light whether
people are merely talking to like-minded people, or if they are actually changing opinion or
behavior. So while recognizing people who talk about your brand could have value, ultimately,
the value of a marketer targeting a high Klout scorer vs. lower Klout score has yet to yield a
concrete business impact.
PeerIndex:16
PeerIndex matches “Peer opinion leaders by providing a service to help
consumers benefit from their concentrated passions, interests and influence particularly with
respect to the services they get from companies; and for brands and agencies to better identify
effective and interested word-of-mouth advocates, that they call “Influencers.”17
PeerIndex
allows users to create their own profile by linking your Facebook and Twitter feeds, but also
“crawls” the Internet and assembles profiles they want to track. Similar to the inputs of Klout,
the site looks at Twitter, Facebook, LinkedIn, Quora, YouTube, and Blog RSS.
Also similar to Klout, PeerIndex scores are on a scale of 1-100. The metric considers
three areas that contribute to a user‟s score. First, “authority” is meant to measure trust and
how much others rely on a user‟s recommendations and opinions on both general and specific
topics. Second, “audience” takes into account how receptive a user‟s audience is as well as the
size of your audience with respect the size of others‟ audiences. Third, “activity/topic
resonance” measures how your actions within a topic community resonate within that same
community. The company also uses their own proprietary algorithm to calculate scores, which
are updated a number of times per week. Peer influence is another important element of the
PeerIndex score. How many postings are retweeted or shared by others as well as who
specifically is replying to or sharing the content can affect a user‟s score. Also like Klout,
PeerIndex specifically suggests ways users can increase their score (i) have a presence on

multiple social media sites, (ii) come up with and share original content with regularity, (iii) build
an active audience and network by engaging in conversations (iv) focus on specific topics and
conversations continuously, and (v) invite friends to join PeerIndex so that PI can track how they
respond to a user. PeerIndex also awards giveaways and special deals, known as “PeerPerks,”
to key influencers on the site. Whereas PeerIndex measures influence by topics across a global
scale, critics of PeerIndex note that the site does not account for list development, influencer
relationship management, campaign management, or engagement metrics, which ultimately
puts the “scores” into question in terms of their utility for a marketer.18
Twitalyzer:19
Twitalyzer measures influence of individuals and brands on Twitter. It is
centered on measuring individual user accounts (as opposed to measuring/monitoring "topics")
and tracks a mix of "measures" and "metrics" over time. Twitalyzer appeals to smaller
businesses who look to the site to identify and provide analytics on top 25 influencers in a given
topic area. All scores reflect a 30-day average, refreshing at least once a day. Different from
other social media influence measures, Twitalyzer scores don‟t use an algorithm. Instead,
scores are calculated using straightforward math to determine the following ten benchmarks for
ranking which are given as percentages: impact, engagement, influence, velocity, generosity,
signal, followers, following, lists, and clout. Here, clout differs from “Klout”, where clout is
defined as the likelihood that you are to be mentioned in Twitter streams of conversation.
However, Eric Peterson, founder of Twitalyzer says their “special sauce” is in the weightings of
these 10 scores. As a way to think about these weightings, Twitalyzer‟s integration with Google
analytics offers a means to take these scores and determine which tweets drove clicks and
sales.
Twitalyzer also measures 5 facets of an individual or brand‟s Twitter activity: influence,
signal, generosity, velocity, and clout. In this way, your peers or the social media population at
large can affect a user‟s score in terms of how many followers a user has and well as how many

of a user‟s postings are retweeted by others. In addition to using Twitter, Twitalyzer also pulls in
data from other social measurement services such as PeerIndex and Klout, despite not having
access to their analytics; they do pool their scores. 20
Many people find the visible calculation of
measurement tools (vs. an algorithm) very refreshing. Further, the ability to identify who the
most influential members of your network are, and when they are online, can be very helpful to
timing a marketing message. That said, for a marketer, the first step is choosing which of
Twitalyzer‟s various measures and metrics tie the most closely to a marketer‟s specific
objectives for being on Twitter (“increase brand awareness” versus “increase customer loyalty
by deepening consumer engagement).”21
The second step is ensuring that the necessary
stakeholders in the Twitter effort buy into them as valid indicators of performance.
Kred:22
Kred identifies the “most influential” people in interest-based communities. It
boasts a fully transparent scoring system of identifying and quantifying influence. A Kred score
is generated by observing user's content, who it reaches, who acts upon it, and whether the
user relays the content of others. The first portion of the score measures a person's “influence:”
their ability to inspire action via retweets, replies, new follows on Twitter, and to generate likes,
comments, shares, mentions, and event invitations on Facebook.
Kred‟s influence score is made up of a normalized 1000-point scale with higher scores
representing higher degrees of trust. Kred‟s second measure considers “outreach:” which
seeks to measure generosity, rewarding action and engagement through spreading others'
information on Twitter via retweets, @replies, and mentions and on Facebook via likes,
comments, and mentions. Kred Outreach Score measured in ever increasing levels and never
decreases. Both areas of scoring provide complete transparency, where users can always see
how actions count towards their scores as well as the scores of others, in real time. Each tweet
has the ability to change your score. At the home page, users are given an "activity statement"

that shows all Twitter and Facebook activity and the points earned through each action.
Retweets, replies, likes, comments, new followers, and listed followers all grow Kred.23
Users can join affinity communities and sharing content with users within your
community typically leads to an increase in both influence and outreach. Users may also give
"+Kred" to others. Moreover, interactions with users of higher Kred can increase a user's score
as well. Real life awards and accolades can also boost your influence level, as they are
evidence that you are a credible source on the subject.24
Many users applaud the truly
transparent nature of the model, as well as, its inclusion of awards and other merits earned in
“real life.” Users also like that they are able to see which communities they are most influential
in, and which specific social media actions they earned Kred points. However, others are
critical that it only measures activity on Twitter and Facebook, but doesn‟t include other metrics
from other social network platforms.
PROskore:25
PROskore, unlike Klout, places emphasis on business-related reputation.
Here again, PROskore is based on a scale of 1-100, where a score is meant to reflect a user‟s
reputation in his/her industry as well as in his/her community at large. In addition to
consideration of work experience and business contacts, the scoring algorithm integrates social
media measurement tools including: LinkedIn popularity and recommendations, Facebook Fan
Pages, as opposed to Facebook profiles, Twitter, Google+, Klout, YouTube, and blog activity.
The company does not share how often a score is calculated. “As professionals are scored and
use the network...we're able to quantify 'who they know'...which has the ability to measure
exactly what their influence really means,” says CEO and Co-Founder Bill Jula.26
“Influencing a
bunch of other Influencers means something different than simply influencing a large number of
people who have no influence themselves.”27

Additional reports allow users to sort a member by location and industry and also
provide access to a variety of marketing and promotional tools, such as email marketing, press
release creation and distribution. These tools help members bump up their scores as a means
of engagement within the PROskore network. PROskore also has an incentive program for
users called PROmos. However, different from other incentive programs that typically involve
free goodies, discounts, and access to exclusive events, PROmos are targeted towards
“professional goals.” As such, high-scoring PROskore users could be invited to a three-month
trial of unlimited Dropbox storage or a voucher to an airlines club lounge in an airport.28
PROskore is more efficient for business use, helping professional entrepreneurs identify new
opportunities and network more efficiently, which sets it apart from Klout which seeks to
measure social influence and automatically imports all personal social media measuring general
popularity or “influence.”29
Thus, in PROskore, users must manually enter their data, which
keeps personal Facebook accounts separate from professional ones. Criticisms of the model
are specific to its limited size. Since not everyone is automatically part of the service, not
everyone will have a ranking. Furthermore, users “skores” won‟t benefit them in the larger
business community that doesn‟t subscribe, because their “skores” aren‟t visible and thus can‟t
be used to compare to competitors.
KEY INFLUENCERS
We now evaluate how adept the online services discussed above are in reporting digital
influence measures that reflect an individual‟s ability to reach and impact his/her target audience
as well as the population at large. In order to do so, we conducted a series of observations
comparing records of several key SMIs across the five leading online services for measuring
digital influence: PeerIndex, Klout, Kred, Twitalyzer, and PROskore. Having thoroughly
examined their metrics and scoring systems, we intend to further analyze their means of

differentiation across their ranking models and their suitability for identifying social media
influence within different sorts of networks.
Subject Pool (SMI)
The observational subject pool includes a diverse group of socially prominent individuals:
Justin Bieber – Hit Teen-Pop Artist
Barack Obama – President of the U.S.
Kim Kardashian – Pop Culture Personality
Oprah Winfrey – Multimedia Mogul/Lifestyle Opinion Leader
Chelsea Handler – Late Night Talk Show Host
Richard Branson – Entrepreneur
Deepak Chopra – Alternative Medicine Proponent
Michelle Phan – Self-made Internet Sensation/YouTube Makeup Tutorial Star
We selected a diverse set of popular individuals to evaluate how representative their real
world influence reach is within their respective affinity networks and the population at large, in
comparison to their digital influence scores. First, each subject in the group had to have an
active presence on each of the digital social media platforms used by the ranking systems
profiled above. For example, each of the five metrics strongly relies on Twitter so it is imperative
for each SMI to have a strong Twitter presence. Further, although individuals may not have
personally registers on the site, PROskore occasionally publishes ranking data on high profile
individuals to demonstrate the aptitude of its metrics in an easy to relate to fashion. With that
restriction in mind, we selected Barack Obama, Oprah Winfrey, Richard Branson, and Justin
Bieber. We then included Kim Kardashian, to observe whether her ranking validates the high
price tag she charges per tweet. We added Michelle Phan to examine how an individual with an
otherwise substantial, yet relatively small constituency within the context of this subject set
would rank. To complement the diversity of the subject pool, Chelsea Handler and Deepak
Chopra were included.
(Subjects‟ social media platforms‟ stats are available in Exhibit 1)

Significance of Metrics Sub-Ranking Analytics
Aside from an overall score ranking, PeerIndex, Klout, and Kred provide a deeper
analytical look at subjects‟ digital impact. Of the sub-ranking analytics available in these
measures, we will discuss the PeerIndex measures of SMI‟s “topics of influence” and “topics of
interest” and Klout‟s tracking of SMI‟s top “topics of authority” as they can be applied to analyze
subject‟s level of influence within particular topics of interest. We will then consider the
significance of PeerIndex and Klout‟s similar ranking measures of SMI‟s respective network‟s
influence, or the aggregate influence of an SMI‟s subsequent “influencees”. Additionally, we will
examine the relationship between Klout‟s measure of “true reach” and “amplification” as it
defines a subject‟s influence density, the ratio of probable level influence within a designated
network. Notably, we concluded that Kred‟s “outreach” measure is not relevant to the scope of
our study. This is because the high popularity and massive number of followers of our pool of
subjects grossly offsets the ratio of the subjects‟ Twitter outreach to other users, deeming this
measure irrelevant.
Procedure
Based on our understanding of the measurement systems and the basic social media
statistics of the SMIs in question (Exhibit 1), we will propose possible results of selective SMI‟s
rankings across measures, as well as, put-forward potential outcomes rankings of subjects
within particular metrics system. We will then draw insights based on the comparison of our
hypothesis with the ultimate findings (Exhibits 2 and 3).
Since Justin Bieber is known to have the greatest number of followers on Twitter and is
famous for his exemplary Klout score, we expect him to rate highest compared to all other SMIs
across all measures except for the PROskore metric. However, as observed in Exhibit 2,
compared to our pool of subjects, Justin Bieber only excels in Klout and Kred rankings. This
finding suggests he may be a more powerful influencer as defined by “reach,” a measurement

element most prominent in these two metrics, rather than “quality of impact” that would propel
his messaging beyond a passive readership.
Having examined Justin Bieber‟s ranking, we move forward to see whether another pop
celebrity known specifically as a major driver of Twitter followers, Kim Kardashian, would also
score highly on Klout and Kred as opposed to the other metrics. Kim has been commissioned
by PR companies to tweet promotional messages on behalf of their clients. Due to the implicit
high market demand for Kim‟s tweets, we expect her to score especially high on Klout and Kred
metrics – which she does (Exhibit 2).
With the election year in progress, and Barack Obama‟s reputation as the “social media
savvy” president, we expect his scores will not be low across any of the metrics. Due to the fact
that his motivation in the social space is to promote his issues and sway constituents to his
convictions, we expect him to outstrip Bieber in PeerIndex. As a politician Obama has a
compelling incentive to interact with his supporters and potential voters. As an entertainer,
Bieber has fans who are important to him, but he does not need to keep a consistent dialogue
with them. Twitalyzer takes an individual‟s ability to drive and maintain conversations into high
consideration of its score. So, rather than focusing on the ability to drive mentions and retweets
alone, we expect Obama to score higher than Bieber. Both hypotheses were realized in the
final score results, as illustrated in Exhibit 2.
Due to PROskore‟s unique emphasis on professional and academic attributes in their
calculation of influence ranking, we expect those SMIs who have PROskore data available, to
rank significantly higher strictly based on those unique emphasized measures, while all other
social media statistics (i.e. Twitter and Facebook followers) will be supplemental rather than
game-changing measures. Consequently, the result confirmed that Obama ranks highest, as
was expected due to his high profile job and prominent education. Oprah and Bieber ranked
evenly, possibly because neither come from a particularly impressive academic background.
PROskore considers Oprah‟s connections to “professionals” to be of high value, while Bieber‟s

sheer “clout” over the Internet manages to even out the playing field. Richard Branson, an
entrepreneur and prolific business owner, ranks above the latter but still falls short of Obama.
We can only assume this is because Obama‟s academic degrees come from higher caliber
institutions, and as president, Obama is in contact with more professionally influential people.
All observed SMIs scored in the 99th
-100th
percentile of the Twitalyzer index, mostly due
to their tremendous presence on Twitter. As such, we expect subjects to score relatively high
and in close range to each other on their Twitalyzer Impact Score. However, this is not the case.
The “impact score” was the only category where Deepak Chopra‟s score excelled above all the
others and Kim Kardashian fell to last in the consideration set. This finding can be attributed to
Twitalyzer‟s emphasis on “generosity” and “responsiveness‟ as a significant contributing factor
to the overall ranking. This indicates that the SMIs in question are not regularly able to engage
with their followers in conversation or share second hand information.
Additional Quantitative Findings
We purposefully selected individuals for our subject pool who would prove to be experts
in similar sets of topics, in order to analyze their impact across comparable categories. Exhibit 5
demonstrates how PeerIndex‟s quantitative analysis of SMI‟s dominance in specific topics can
be used to go beyond general influence rankings and particularly identify influencers per field of
interest.
Both PeerIndex and Klout measure the influence score of one‟s pool of” influencees.”
PeerIndex displays the influence scores that rank “influencees”. Klout provides an aggregate
“network impact measure”. It is interesting to observe the relationship between an SMI‟s ranking
on either of these services versus their respective influencees‟ rankings (average of top ten for
PeerIndex). In observing Exhibit 4, we learn that the SMIs‟ Klout and PeerIndex scores do not
correlate with their respective networks‟ influence scores. We therefore assume the magnitude
of influence of Deepak Chopra is likely to be greater than the magnitude of influence of Justin
Bieber. This is because Deepak Chopra is more likely to be connected to individuals of higher

influence. Understanding the relative magnitude of influence for a given individual can be useful
for marketers to effectively maximize the spread of messages and action-taking behaviors. For
instance, while Chelsea Handler ranks relatively high by both measures, her pool of influencees
are not comparatively impactful; so while Chelsea ranks as relatively impactful, her messages
lack an effective longevity and may not go on to impact other two-degree removed followers.
The Klout score takes a number of factors into account that may obscure SMIs‟ relative
ability to effectively influence those in their networks of interest. When companies look to
contract celebrities for endorsement, they often overestimate the significance of a high Klout
score. Rather, they should take into greater account the ability to evoke action by others. The
“reach/amplification” ratio (Exhibit 6) offers an excellent alternative to examine how many
people are likely to be influenced by an SMI‟s single action or message. In this case, “true
reach,” which measures how many individuals an SMI‟s message is able to reach per day and
“amplification,” which measures the density of impact on a particular network. Here we see that
even though Michelle Phan falls in the middle of the comparative score range, she has a
phenomenal ability to influence her relatively small group of niche followers. Thus, when it
comes to examining the degree of impact versus reach, we require further analysis as to the
drivers behind each metric‟s respective scores.
Additional Qualitative Findings
PeerIndex has two unique methods of tracking and ranking which measure: (i) the top
trending topics that an SMI chooses to focus on and (ii) comments posted by other users that
mention the particular SMI (Exhibit 3). These measures are by no means perfect, as they use
embedded links, such as retweets and word recognitions to rank the top “category” related
words that other users most often use to associate with either a specific SMI or the top topics
generated by that SMI upon which users are most likely to take note or take action. For
instance, Justin Bieber indexes high on messaging about music and is often associated with
A.M.E. (arts/music/entertainment) as well as life/leisure. These are very similar topics, as the

former folds into the broader category of the latter. SMIs hope to be impactful about the topics
upon which they choose to comment, however, these are not always necessarily the topics over
which they have the most impact. For instance, while Obama is heavily messaging about
taxation, veterans, and automotive and manufacturing industries, it is important to note that
those specific topics are not resonating with the broader digital community. Instead, he is most
influential in the broader areas of politics, health, business, economics, science, and medicine.
These topics are natural for a president to be deemed influential. However, a longitudinal study
with a deeper analysis of influence drivers may reconcile the discrepancy between Obama‟s
goal message and his perceived message.
Similar to PeerIndex, Klout also tracks the top topics of “authority” for each SMI (Exhibit
7). Under closer observation, this measure is highly flawed for several reasons. While not fully
disclosed, it is evident that Klout‟s means of tracking “authority” topics entail a broad web survey
of all discussions that mention an individual‟s name or associated Twitter handle. Additionally,
the measure utilizes a qualitative rather than a quantitative scale, ranging from high to weak.
This inherently presents challenges for analysis. This method of tracking often leads to
misleading or subjective observations. For instance, Justin Bieber, by popular association, is
tracking as a “strong” authority on topics of “music” and “Justin Bieber.” However, it is surprising
to find that Bieber is also a strong authority on the “Holocaust” and “Adolf Hitler” tracking as
“high” right behind those categories. A quick Google search reveals a rampant rumor claiming
that Justin Bieber openly denied the Holocaust. Whether or not this is true, Klout picks up on all
mentions where Bieber is associated with the Holocaust, and as such he is inevitably identified
as an “authority” on the topic. Similarly, we uncovered that Barack Obama tracks as a “high”
authority on home brewing because he recently had local home-brewed beer served at The
White House Correspondents‟ Association Dinner. Kim Kardashian is found to be of “medium”
authority on Harvard Business School due to recent rumors hinting at her romantic interest in
Jeremy Lin, who graduated from Harvard‟s undergraduate program in 2010.

DISCUSSION
For decades, celebrities have been hired to endorse products and services ranging from
Wheaties to Weight Watchers. With the advent of social media, in particular the popularity of
Twitter, the latest trend is to pay celebrities to tweet about your brand or cause. Kim
Kardashian is allegedly paid $10,000 per tweet, which speaks to how powerful her brand must
believe in influencing consumers. It is not an easy feat to quantify influence in social networks,
but online social networks have devised their own metrics to track activity and thus, many social
scientists have run studies to determine how information is diffused through networks.
Presumably, high profile public figures have large fan bases and are therefore, viewed as
having influence over a large group of people. However, it is not certain that fans or followers
are actually listening to what these so-called influencers are posting or tweeting about and/or
causing followers to change their behaviors accordingly. Whereas Chelsea Handler has over
twice as many Twitter followers than does Richard Branson, Branson‟s posts have been
retweeted twice as many times as those of Handler‟s have been over the past ninety days.
Although high profile public figures have large numbers of followers, it appears that the
structure of their networks matters more than the quantity of individuals within their networks.
According to Watts and Dodds, “the ability of any individual to trigger a cascade depends much
more on the global structure of the influence network than on his or her personal degree of
influence – that is, if the network permits global cascades, virtually anyone can start one, and if
it does not permit global cascades, nobody can.”30
They found that the size of the cascade
depends on the average density of the network to be influenced. When the average density is
too low, there are not many edges between nodes and so information is prevented from
spreading to many other nodes. Conversely, when the average density is too high, nodes are
connected to many other nodes and will only adopt the behavior if several of its local neighbors
have already adopted it.31

Similarly, Aral, Muchnik, and Sundararajan claim that “cohesive, dense local networks
(with more ties among their friends) adopt at a higher rate in the presence of an adopter friend
controlling for observed homophily, reinforcing prior arguments that cohesive networks magnify
information exchange and persuasion via redundancy and trust.”32
Aral et al. observed the
adoption of a mobile application within a network of twenty-seven million individuals to
distinguish whether adoption spread as a result of peer-to-peer influence or due to homophily.
Both types of diffusion are correlated with network structure, particularly in relation to assortative
mixing and temporal clustering.33
Nodes may be influencing other nodes that they are
connected to and thus transmitting information or behavior through the network. However, the
case may be that nodes are adopting the same behavior because they are characteristically
similar to and share the same interests as one another (e.g. Michelle Phan has an exceptionally
devoted following of women interested in cosmetics application techniques. We find she also
has a high Klout amplification score, indicating a strong probability that any one of her tweets or
posts will incite action, such as a retweet or a comment). After the five-month period following
the launch of the mobile service, the evidence pointed out that friends were up to five times
more likely to adopt the service within two days of their friends‟ adopting it.34
This suggests that
peer influence is at play but it is not a definitive conclusion. It also appears that homophily may
be mistaken for influence-based contagion in the early stages of diffusion.35
This raises a question about the influencer‟s audience and their likelihood to adopt new
behavior as a result of observing the influencer promoting the behavior. The success of
cascades may be contingent upon the presence of a large number of easily influenced
individuals.36
For instance, the majority of Justin‟s Bieber‟s fans consist of young girls who may
have crushes on him or idolize him. Therefore, the fans are more likely to retweet any comment
that Justin Bieber makes. In contrast, Oprah Winfrey‟s fans may engage with her content on a
more topic specific basis versus comment on every one of her tweets or postings. Watts and

Dodds seem to have mixed results for their “influentials hypothesis” as they found that anyone,
deemed influential or not, can spur a cascade; that said, they found that the largest cascades
were most often initiated by influentials.37
Despite the detection of cascades, Bakshy, Hofman,
Mason, and Watts claim almost all are small and shallow; only a tiny fraction of cascades reach
thousands of individuals.38
Another study proposes that high profile figures are more likely to propel a successful
cascade when they act as intermediaries as opposed to the originators of the behavior or
information.39
Budak, Agrawal, and Abbadi claim there are four different influencer identities
present in the blogosphere and who cause successful cascades of information diffusion
because they span network structural holes. Three out of the four types of influencers are taken
from Malcolm Gladwell‟s “The Tipping Point.” Budak et al. characterizes the four types as
follows: Connectors are individuals who hold central positions within the network; mavens are
original sources of many new information cascades, who have significant influence over their
neighbors; salesmen are individuals who are persistent in activating cascades; and translators
leverage their multi community memberships by listening to one community and then restating
the content in a version that is applicable to a different community.40
The results of the study
suggested that the most expansive reach is through a combination of these influencer types.
For instance, starting with a maven and then moving to a connector results in the greatest
magnitude of the cascade and next most successful cascade includes salesmen and
translators.41
As discussed in the case studies above, Budak et al. believe “that different social
networks provide different ways of interacting which means that certain actors, while not so
significant in certain networks, can be highly influential in others.”42

CONCLUSION
This paper examined the various measures of influence in social media. We evaluated
the strengths and limitations of the leading models by comparing them first to one another and
then as a means to better comprehend top-ranking key influencers. We examined the potential
causes for disparate rankings across the different metrics. In doing so, we observed the
inherent attention that the SMIs receive, but we were unable to precisely identify their true
influence and ability to change behavior. We then observed some of the contributing
components to the “title metrics,” such as “amplification,” “true reach,” and “network impact.” It
is in using these components rather than the “title metrics” which may make us better able to
predict the relative magnitude of reach and impact of a particular influencer‟s message.
We also examined the presence of homophily and external environmental factors in
social media platforms and concluded that they played more of a role in confirming correlation
across the networks than did the presence of influence or causation. The Shuffle and Edge
Reversal Tests illustrate that original and time-shuffled/edge reversed tags follow the same
path, indicating a lack of definitive influence. Moreover, in Aral‟s study of mobile application
adoption, the authors found that previous methods grossly overestimate peer influence in
product adoption decisions in networks and find that homophily explains more than 50% of the
perceived behavioral contagion.43
Thus, while we are certainly able to observe correlation that
is likely due to homophily, we still cannot discount the presence of influence, which would
suggest causation.
We posit that further research must be done to determine the drivers of a given SMI‟s
ability to incite action and change behavior in social media. We recommend studies that are
able to isolate homophily and focus determining whether the key drivers in inciting changes in
behavior are due to (i) peer-to-peer influence, (ii) external influence from beyond an individual‟s
network, or (iii) due to some as yet to be determined cause.

Networks, Crowds & Markets Final Paper

REFERENCES
1
Ceren Budak, Divyakant Agrawal, and Amr El Abbadi. Where The Blogs Tip: Connectors,
Mavens, Salesmen, and Translators of the Blogosphere. SIGKDD Workshop on Social Media
Analytics, 2010.
2
Ceren Budak, Divyakant Agrawal, and Amr El Abbadi. Structural Trend Analysis for Online
Social Networks. UCSB, 2011.
3
4
5
A. Anagnostopoulos, R. Kumar, M. Mahdian, Influence and Correlation in Social Networks, 1-9,
2008.
6
2008.
7
2008.
8
N.A. Christakis and J.H Fowler, The Spread of Obesity in a Large Social Network Over 32 Years,
The New England Journal of Medicine, 357 (4): 370-379, 2007.
9
2008.
10
A. Anagnostopoulos, R. Kumar, M. Mahdian, Influence and Correlation in Social Networks, 1-
9, 2008.
11
9, 2008.
12
9, 2008.
13
Klout Homepage. http://klout.com/corp/about
14
Ibid.
15
BrandSavant Blog. http://brandsavant.com/on-klout-bashing/
16
PeerIndex Homepage. http://www.peerindex.com/help/scores
17
Ibid.
18
Ibid.
19
BlogTalkRadio Website
http://www.blogtalkradio.com/breakthroughbusiness/2011/04/25/social-media-metrics-
twitalyzer-bbsradio-michele-price
20
Twitalyzer Homepage. http://twitalyzer.com/
21
Gillion On Data Blog. http://www.gilliganondata.com/index.php/2010/11/09/Twitter-
performance-measurement-
withtwitalyzer/?utm_source=quora&utm_medium=question_response&utm_campaign=what_i
s_twitalyzer
22
Kred Homepage, slideshow. http://kred.com/
23
Ibid.
24
Ibid.

25
PROskore Homepage Video. http://proskore.com/
26
Digital Journal article. http://digitaljournal.com/article/314817#ixzz1u2L8bQWP
27
PROskore Homepage Video. http://proskore.com/
28
Ibid.
29
Ibid.
30
Duncan J. Watts and Peter Sheridan Dodds. Influentials, Networks, and Public Opinion
Formation. Journal of Consumer Research, 2007.
31
32
Sinan Aral, Lev Muchnik, and Arun Sundararajan. Distinguishing Influence-Based Contagion
From Homophily-Driven Diffusion In Dynamic Networks. NYU Stern, 2009.
33
34
35
From Homophily Driven Diffusion In Dynamic Networks. NYU Stern, 2009.
36
37
38
Eytan Bakshy, Jake M. Hofman, Winter A. Mason, and Duncan J. Watts. Identifying
„Influencers‟ on Twitter. WSDM, 2011.
39
Analytics, 2010.
40
Analytics, 2010.
41
Analytics, 2010.
42
Analytics, 2010.
43

Networks, Crowds & Markets Final Paper

More Related Content

What's hot (13)

Viewers also liked (8)

Similar to Networks, Crowds & Markets Final Paper (20)

More from Jason Rawlins (7)

Recently uploaded (20)

Networks, Crowds & Markets Final Paper